This paper presents a learning algorithm for neural networks having general feedback connections. Simulation results are presented to compare the algorithm with the existing backpropagation algorithm; the former is shown to have a greater capability in realizing complex tasks.